Apple Silicon Predictions

LuxTerra

Limp Gawd
Joined
Jul 3, 2017
Messages
218
No insider info here, just looking to have some fun.

BLUF: Apple's product differentiation strategy is different than the Intel nickel and dime you to death for ever 100Mhz or feature. We're going to see a more unified product stack, consistent with their mobile strategy, that's differentiated by form factor and quantity of cores/RAM/storage. Optimizing their lineup, Apple will use SoCs for all mobile (M series) products and will be 100% integrated GPUs until die/power constraints force them to an architecture variant with discrete CPU/GPU (although multi-chip packaging is possible) for the top tier products like the Mac Pro. Assuming "traditional "naming convention, Apple will have three SoCs: e.g. M1, M1X, and M1Z; the "TDP" targets will be 25W, 50W, and 75+W respectively. To minimize design costs, these will likely not cross over in many products. Apple will tier their offerings by die harvesting core counts, but not any other features or clock speeds. Specifically, this means (or the updated equivalent of):

16" MBP (M1Z at 75+W): 12+4c CPU, 24c GPU, up to 64GB LPDDR4 with double the M1 channels and double the die stacks to meet this. Die harvesting for tiered product lineup would look something like 10+4c CPU and 16/20c GPU options. Yields will dictate where the line is drawn. Die size should be approx. 250mm^2.​
14" MBP (M1X at 50W) 8+4c CPU, 16c GPU, up to 32GB LPDDR4 with the same channel width as M1. Die harvesting for a tired product lineup would look something like a 6+4c CPU and 12c GPU. Die size should be approx. 180mm^2.​
13" MBP entry level is unlikely to be updated and may follow the iPhone SE approach of using old stock/designs. Will stick with the ~25W M1 at 4+4c/8c.​



Discusison
Key assumptions include:
  1. Apple will remain consistent to their core strategy of focusing on efficient compute and differentiation by form factor and simple logical jumps in quantity (e.g. core counts). In a sense, this really is setting manufacturable die size mm^2 and power targets and selling the most that fits in that envelope. I'm guess those are:
    1. 100-120mm^2 / 25W
    2. 150-180mm^2 / 50W
    3. 200-250mm^2 / 75W+
  2. Apple will continue to be and pay for the bleeding edge of TSMC nodes necessitating a fair amount of die harvesting as die get larger.
  3. Apple will optimize their lineup by limiting the number of SoCs or discrete chips they need to manufacture.
  4. Apple will continue to "thin out" their form factors to optimize mobility.
I think those are reasonable and consistent with Apple's strategy.

Form factors will need an update to bring new tech (e.g. possibly miniLED, faceID, etc) and to keep the look-at-me-I-have-the-newest-stuff fan boys satisfied. The 16" is on the thicker larger side of things for an Apple mobile product and I expect them to try and thin it out to hit a ~4lb weight target while being sleeker. Their traditional upper end TDP target has been an "official" ~45W CPU and ~50W GPU, but...we all know those TDPs are mostly a joke for Intel and there have been cooling/performance complaints with this product over the years. I expect Apple to target a more nominal 75W TDP with limited 100W peak TDP usage. I don't see Apple hitting their battery life targets with a capped 100W/hr battery capacity unless they reign in the peak power consumption a bit. I really think the current 16" is the chassis they needed to support Intel's products and it's not really the portable desktop chassis Apple wanted. GPU will be a huge factor in selling the 16" over the updated 14" as is possibly more 8k focused accelerators and double the RAM channels to support that many CPU/GPU cores. The only offering here is what I've notionally called the M1Z. The smaller M1X won't be offered because the SoC will nominally required a different motherboard design/integration due to power/RAM configurations.

The rumored 14" is needed to allow TDP to grow towards the 50W range and allow for new tech. I don't expect any new features over the M1 here, just more. Again, only the M1X will be offered to minimize the design/integration costs, at least for now. It remains to be seen how what has been the entry level 13" MBP works out in the long run vs the MBA. Designing and integrating both the M1 and M1X into the 13"/14" MBP might not be a problem though, since I expect them to be just a smaller/larger version of each other. Unlike the M1Z, which I think will necessitate more RAM channels, etc. to feed the beast.

Now, for anything larger/discrete, Apple will have a separate lineup. Let's call it the P1 and P1X/Z. Beyond ~250mm^2 yields on a bleeding edge process are just not ideal. Plus, for max performance, Apple will want to wait a year for TSMCs performance version of any node. It's not that you can't build these, but rather that it's quickly hurts margins and Apples not know for doing that! This likely puts Apple on a 2yr refresh cycle. New nodes bring updated M lineup, optimized nodes the following year bring top end P products. Since the GPU doesn't have to be on the monolithic die for both area and power constraints, the P lineup should easily see 2x core counts over the M lineup. It's reasonable for the very top tier Apple will have 3x core counts just based on die/power estimates. This would match the 32+4c CPU and 128c GPU rumors. Also, I would expect that the very top tier P1X/Z Mac Pros would allow for dual sockets with a very nice profit margin. :)

TSMC 3nm node in late 2022 should allow all M2 products to have 50% more CPU cores and 100% more GPU cores. Apples product requirements seem to suggest that they will prioritize the GPU over more CPU cores.
 

LuxTerra

Limp Gawd
Joined
Jul 3, 2017
Messages
218
No one wants to play market analyst? :)

Well, I'll take the time to address a counter point. There's several good arguments that my estimate is far to aggressive and that the top end M1 variant will be the 12x core 8+4c variant that's "leaked." My thoughts on why that's a reasonable statement are:

  • This is gen 1 and Apple doesn't have to knock it out of the park. Merely being competitive is likely good enough. The current 16" Intel MBP tops out at 8x cores, so why go further?
  • Even this 8+4c M1 variant is still a large chip for Apple. Probably around the 180mm^2 die size. Apples on a bleeding edge process, so why take risks?
  • 8+4c CPU can still be die harvested into a 6c+4 mid-range product, why build three SoCs?
  • The 8+4c is an alleged leak from someone who actually used it, why doubt it?
  • So far, M1 only die harvests the GPU, not CPU cores.

My counter points are:

  • We have lots of historical indications that Apple's never liked Intel's slow roll, forced by manufacturing processes or not. This whole transition hinges on Apple being able to enable the "experience" they want and that's recently been far more than Intel could deliver.

  • Relevant to the previous point, Apple has already shown that they're willing to be aggressive. Namely, they took what would have been a top of the line quad core 13" MBP and shoved it all the ways down into a MBA! That's an aggressive roll out and I don't see them changing things.

  • Historically, Apple's never updated the Mac lineup like they have the iPhone/iPad lineups. There's a lot of good reasons for that, including relatively low sales volume. Plus, Apple has a lot on their plate with this whole 2yr transition. I think they want to be aggressive, but are not likely to take unnecessary risks. Adding a M1->M2 update onto of launching the higher tier products just seem unnecessary risky and doesn't align with TSMCs published roadmaps. This 2yr update cycle implies that Apple needs to not only fend of Zen3/11th Gen, but also Zen 4 and Alder Lake with this launch. This heavily suggests an aggressive strategy so their offering is viable into 2022.

  • Yes, an 8+4c / 6+4c / 4+4c lineup could work, it just seems unreasonably tight and provides limited to no inter-form factor stratification; i.e. a lot of overlap. For example, a 16" then gets the 8+4c and the 6+4c, while the 13/14" gets a 6+4c and 4+4c? I just don't see that happening unless Apple pushes the GPU cores harder than we think. A 12+4c / 8+4c / 4+4c strategy provides a lot more differentiation. Differentiation ONLY by die harvesting just doesn't seem like an Apply strategy IMHO.

  • As the dies get bigger, so to does the likelihood of need to die harvest CPU and GPU cores.

  • The TDPs just don't work out. Each high performance Apple A14/M1 core draws about 4W. Maybe a bit more, maybe a bit less. Each GPU core seems to draw about 1W. All power numbers are at TSMC 5nm since that's the only metric we have. Now, an 8+4c CPU would need 16W more than the M1. Given the M1 is approx. 25W, that puts an 8+4c around 40W TDP. That's a bit much for sustained performance in the current 13", but probably fits within the assumed 14" update(?). Furthermore, the higher end 13" MBP always had a lot more graphics power, but even doubling that only puts this near my estimated 14" TDP target of 50W. Now that works for an updated small MBP, but is way under powered for a 16" MBP which had "fake" TDPs of 45W CPU and 50W GPU. I just don't see Apple taking it easy and not pushing the 16" performance and power levels. Even tripling the GPU keeps the TDP well under 60W and again, that just seems like too little for the 16" MBP targets and market.

  • Whatever goes in the 16" MBP likely goes into the iMac and the peak 8+4c option just seems way too little for a solid desktop.

  • Assuming the 8+4c leak is valid, then Apples been upstaged before release by their own demo, and that's just not like them. The company that brought you and "one more thing" isn't going to let a the top end configuration leak months in advance. They know these leaks will happen and likely sanction them, but why leak your best too early? Makes no sense to me. Yes, I think the M1X will top out at 8+4c, but that's highly unlikely IMHO to be the top end 16" MBP option. We've never seen Apple work that way.

  • If the rumors are true, Apple silicon is far cheaper to Apple than Intel silicon. Now, Apple could just absorb the higher profit margins, but given their existing profit margins, why bother? That's not an absurd statement. Yes, they obvious prefer making more money. However, building your own silicon is hard and expensive; even if you reuse your smartphone/tablet IP. It's highly unlikely that Apple looked at their IP and looked at the market and said, you know ~7-8% market share is fine with us. Apple has a tremendous opportunity here and I think they would be fools to pass up taking full advantage of their leadership and not going all out. This is an ecosystem buildout and the value of the ecosystem is far greater than the sum of the components.

  • Finally, the 12+4c / 24c strategy isn't the most they can do. It's the aggressive strategy, but not the crazy aggressive strategy. Realize that my 75W+ target is far less than the old Intel solutions and that TDP was a joke. If you only ask, what could they technically do, regardless of how mad the MBAs get, then a 12+4c / 24c at ~250mm^2 isn't it. It's unlikely TSMC can push the reticle limit toward the 600+mm^2 yet; at least not at any price anyone will pay. If they could, it's likely NVIDIA would already be there with their datacenter GPUs. Dies that large won't come until the performance 5nm TSMC process later this year at best. However, they could push into the 300-400+mm^2...technically, not pragmatically; so many angry MBAs! Now that SoC could support a 16+4c / 32c option based on the die shots/logic sizes I've seen. That's quadruple an M1 mind you, but not everything needs to be quadrupled. That could keep sizing a bit smaller than 4x M1 die size or free up space for extra accelerators (e.g. for 8k video processing). Now, if we keep the power scaling of M1, that's merely 16x CPU cores at 4W for ~64W total, plus 32 GPU cores for 32W. That actually hits the current 16" MBP official TDPs really well, but as we covered, TDP with Intel is a bad joke. So technically, that's probably the very upper limit currently on a TSMC 5nm node that they could push, but what an absolutely spectacular laptop that would be! Consider that we'd be talking about 16x fast CPU cores, where each core matches a Zen3! Consider that the GPU would be 4x an M1 and an M1 is already about a GTX1650...multiply that by four and let me know what you think. All of that for 100W sustained would be nuts. Now, I have no reasonable belief Apple has let the engineers go insane and create such a thing. At least not for a 16" MBP. It will come, but likely not in mobile form factors. One can dream though.
 
Last edited:

paradoxical

Weaksauce
Joined
May 5, 2013
Messages
102
I actually agree with most of what you wrote in the OP. I will personally be in the market for a loaded 8 core 14in. My dream spec is 8+4 core CPU, 8-16c gpu, 4TB SSD, 32GB RAM, 4+ thunderbolt ports at full speed. I'd pay $3k for that if it was MacOS only, and $5k for it if it could emulate windows.
 

UnknownSouljer

Supreme [H]ardness
Joined
Sep 24, 2001
Messages
6,957
I think a more interesting conversation to have at this point isn't what the CPU's are going to look like but more what the GPU end is going to look like (although obviously this stuff is mentioned in the op). I say this mostly because I think most can see that even if the only thing Apple does this year is increase core count, the CPU side of their design is already destroying what AMD and Intel are capable of doing.

It's uncertain whether or not Apple will continue its partnership with AMD for GPU's or not. And it's kind of hard to believe that Apple will have a GPU in two years that is faster than 4x Vega II's as found in the maximum configuration in the Mac Pro. And that is putting aside any advancements that AMD of course is already doing with RDNA2 which of course furthers that GPU gap.

It's less of a leap for Apple to be able to beat the 5500M that's in the current 16" MBPt, but honestly even that is pushing it. Unless Apple has a wild card up their sleeve and they've been developing discrete GPU tech all along - it's hard to think that they have a surprise that will put their laptop squarely against the other workstation monster GPU's in the space. If Apple actually wants to increase in market share it will have to have systems that can go up against mobile workstation GPU's, which at least up to this point they have been unwilling to do. Mostly because they have always wanted a balanced machine that actually has usable portable battery life and fits a very particular smaller form factor. However with their CPU's not taking up a lot of wattage/heat, they could put in a much more massive professional discrete GPU from AMD without the same penalties that they would have if they were still on Intel CPUs. Or like I say they can alternatively show us they have some GPU monster that they've been hiding. Because I don't think that simply increasing GPU core count on M1 will be enough.
 
Last edited:

LuxTerra

Limp Gawd
Joined
Jul 3, 2017
Messages
218
I think a more interesting conversation to have at this point isn't what the CPU's are going to look like but more what the GPU end is going to look like (although obviously this stuff is mentioned in the op). I say this mostly because I think most can see that even if the only thing Apple does this year is increase core count, the CPU side of their design is already destroying what AMD and Intel are capable of doing.

It's uncertain whether or not Apple will continue its partnership with AMD for GPU's or not. And it's kind of hard to believe that Apple will have a GPU in two years that is faster than 4x Vega II's as found in the maximum configuration in the Mac Pro. And that is putting aside any advancements that AMD of course is already doing with RDNA2 which of course furthers that GPU gap.

It's less of a leap for AMD to be able to beat the 5500M that's in the current 16" MBPt, buthonestly even that is pushing it. Unless Apple has a wild card up their sleeve and they've been developing discrete GPU tech all along - it's hard to think that they have a surprise that will put their laptop squarely against the other workstation monster GPU's in the space. If Apple actually wants to increase in market share it will have to have systems that can go up against mobile workstation GPU's, which at least up to this point they have been unwilling to do. Mostly because they have always wanted a balanced machine that actually has usable portable battery life and fits a very particular smaller form factor. However with their CPU's not taking up a lot of wattage/heat, they could put in a much more massive professional discrete GPU from AMD without the same penalties that they would have if they were still on Intel CPUs. Or like I say they can alternatively show us they have some GPU monster that they've been hiding. Because I don't think that simply increasing GPU core count on M1 will be enough.
The GPUs are extremely interesting besides core counts. There's some reverse engineering going and and as far as we can tell, it looks like they built a very scalable, but simple GPU architecture. Meaning Apple's GPU isn't going to put the hurt on Nvidia's compute business anytime soon, but as a purely graphics accelerator, it's crazy efficient. See https://asahilinux.org and https://rosenzweig.io/blog/asahi-gpu-part-1.html

That's what makes the M1 GPU extremely interesting. It looks absolutely amazing for power efficiency compared to traditional Nvidia/AMD GPUs. It seems to be approx. 1-1.25W/core and an 8x core GPU is beating a GTX 1050Ti in most benchmarks and comes up a bit short of the GTX 1650. Given that the M1 8c GPU is claimed to be 2.6TFLOPS and the GTX 1650 is 2.9TFLOPS, Apple seems to be about as efficient per theoretical TFLOP as Nvidia. I'm going to use Nvidia as the baseline since they currently have the top dog with the RTX3090.

Let's estimate it two ways. First by TFLOP and second by a simplistic GPU benchmark score. By TFLOPS, in comparison to the RTX 3090, which is rated at 36.6 F32 TFLOPS, ignoring the RT/Tensor cores and focusing only the shaders floating point performance. Now, it's not a fair comparison since those RT/T cores can provide other value and certainly take up a lot of die space, but we're playing FLOPS and hand grenades here. That would require Apple to increase the M1 GPU cores by approximately 14x (36.6/2.6); that's a 112 core M1 GPU. The rumors are that Apple is working on a 128c GPU for the Mac Pro, so that's at least in the ball park. Now Nvidia is still on TSMC 7nm and this hypothetical Apple GPU likely won't be seen till 2022. Now, is that reasonable die and power constraints? Yes. At 1.25W/core, that's a mere 192W. Toss in some overhead and we have a ~200-250W GPU. Perfectly reasonable compared to the RTX 3090s 350W.

Based on the A14 4c GPU (https://www.techradar.com/news/appl...outperform-some-amd-and-nvidia-graphics-cards) we can reasonably assume that it's about 3mm^2 per GPU core for the M1. A hypothetical 128c M1 GPU could be ~400mm^2. Now that's currently a huge TSMC 5nm die, but it wouldn't be built on the low power 5nm, but rather the performance variant due later this year. It's also only 2/3rds the size of the 7nm RTX 3090 die (628mm^2). Being fair, the RTX 3090 does have all those RT/T cores and is on the 7P process. It seems that Apple has, at least for the M1/A14 GPUs, optimized for simplicity and low power over features and die area. However, none of those numbers hurt my head or seem unreasonable. Apple's hypothetical 128c GPU likely won't be as fast as what ever Nvidia can make on a TSMC 5P process, nor will it have all the compute cores/features (RT, Tensor, and CUDA), but it should be providing a lot of performance for a good bit less power (7P to 5P isn't going to make an RTX 3090 equivalent drop 100-150W to get near Apples 128c GPU...at best the scaling could bring it parity, but rarely does a real product achieve full TSMC marketing scaling in either power or performance).

The other way we can do a quick napkin check is to look at benchmarks. You can Google and use whatever you prefer, but a rough estimate is that an RTX 3090 delivers about 6-8x the real world performance of the 1050Ti, which an M1 handles easily. This is shader cores to shader cores only comparison; no RT/T. Now an M1 is 8c, so 8x would be only 64c! This implies that for basic graphics Apple has better performance scaling, but we can all agree building a small, 8c GPU is nothing at all like building a 128c monsters; it's likely linear scaling is just wrong. However, it does give a reasonable upper limit to the hypothetical 128c M1 GPU at 2x an RTX 3090 shader cores only. Don't forget, 5nm vs 7nm comparison here.

Using those two ranges and adjusting for the fact Nvidia will be on TSMC 5P by then, it seems reasonable that the best case Apple could hope for with a monster 128c GPU would be to match Nvidia's best in 2022 (RTX 4090?) and will likely come up short on performance (close enough for practical purposes), but at lower power. Not an easy task mind you, but not an insane estimate either. Apple has the engineering expertise and money, but do they have the will to build it? Hard to say, but to date, they've been fairly aggressive. Maybe generation 1 will skip a discrete Apple GPU, but I kinda doubt generation 2 does...the profit margins are too great to ignore.


As for AMD MBP GPUs, run the same ball park estimates and it seems highly unlikely Apple has any reason to put another AMD GPU in a 16" MBP. AFAIK, we don't know exactly how good RDNA 2 mobile solutions will be, but Navi 23 is the rumored 5500M replacement, but with 65-90W TDPs. Navi24 rumors at ~50W TDP are a better fit for Apple's likely TDP targets. Navi 23 is rumored to be 32CUs compared to Navi 21s 80CUs for 22.5TFLOPS. I know, all rumors, but simple scaling says the ball park 65-90W TDP Navi 23 should deliver about 9TFLOPs. That's just over 3x an M1. The rumored 24c M1X (or M1Z in my OP) would roughly cover that performance wise, but it should in theory do so at half to a third the power (24 cores at 1.25W/core peak). Die size at 3mm^2 per GPU core and it's a no brainer. I just don't see RDNA 2 making it in ANY Apple laptop. They burn far more power, for similar performance. It's the same story with the CPU cores.

Now, if you care about various compute capabilities, it's obvious that Apple has some work to do. However, the M1 8c GPU is no OpenCL slouch. It's roughly half a 5500M, so a 24c GPU should compare favorably to a Navi 23/24.

Memory bandwidth is the elephant in the tent here and that's one reason I think they will split into a 3 SoC strategy. They really need to double the LPDDR4 interfaces to go from ~60GB/sec with the M1 to 120GB/sec for my 16" MBP estimates. Gotta feed the beast and even Apple's memory compression techniques can't get around just how much bandwidth a 8-12c CPU and 16-24c GPU will need to compete with Navi 23/24, etc.



Edit: now, that's a lot of speculation based on rumors. Rumors are constantly shifting and hardly ever totally accurate. However, that's why I'm just doing ball park estimates. Who's a bit faster than whom is hard to say, but there are three things I'm fairly certain about. 1) Apple makes far more money selling Apple silicon than Intel or even AMD...this will encourage an aggressive rollout. 2) Apple has a clear power advantage by having a tightly integrate and focused feature set architecture. 3) Apple can roughly compete on performance, but it doesn't matter in the slightest if their only 90% Nvidia/AMD/Intel if #1 and #2 are true. They just have to be close enough for practical purposes.
 
Last edited:

UnknownSouljer

Supreme [H]ardness
Joined
Sep 24, 2001
Messages
6,957
I love Apple, but I think even if things could scale linearly (and we both know they can't) it's optimistic to think they can ramp up to a 128 core GPU in that short a time frame. I'm also not entirely certain if at that point they'll still try to have the GPU and CPU still be packaged into an SOC or not either. As honestly it might be untennable from just a thermals perspective and also limiting in the sense of upgrades. That's less of an issue on a laptop, and more of an issue inside of the Mac Pro. Of course that begs the question too of, what exactly would be user upgrade-able inside of a machine that is basically 100% proprietary - and that is in fact a very interesting question - but in theory if they allowed for CPU/GPU upgrades (even if only directly through Apple themselves) that would/could increase the longevity of the machines and still perhaps keep standardized slots for PCI-E and RAM - which would lower the cost of those components not only just for the user base but also Apple themselves.

As another wrench, I think Apple also wants to be a premiere machine in the graphics space. It's little known, even to Apple fans, but Metal has API's already in it for both RT and VR. Apple launched both of those API's over 3 years ago. They were actually ahead of the curve at least in the sense of ensuring that those things could be coded for on macOS. However they've been lacking the hardware support. Obviously AMD only recently has cards that can support RTX features and there isn't really any game development or VR happening on macOS - but the intention again is to capture that space. Apple is going to have to do a heck ton to catch up there too if they want to attract devs and consumers. Not having RT is going to matter more and more and it's not something Apple can ignore for long. Not even in their first gen product. In 2 years when the Mac Pro launches if it doesn't have RT that 's going to be woefully behind. That will likely be 3 RT generations for nVidia and at least 2 for AMD. To not have it on a professional level machine will just be a non-starter for basically all devs and it won't make sense at all versus just buying and building a WinTel machine.

That's why I think at least short term they're going to have to have an AMD discrete GPU. I'm not sure there is any logical way around it.
 

LuxTerra

Limp Gawd
Joined
Jul 3, 2017
Messages
218
I love Apple, but I think even if things could scale linearly (and we both know they can't) it's optimistic to think they can ramp up to a 128 core GPU in that short a time frame. I'm also not entirely certain if at that point they'll still try to have the GPU and CPU still be packaged into an SOC or not either. As honestly it might be untennable from just a thermals perspective and also limiting in the sense of upgrades. That's less of an issue on a laptop, and more of an issue inside of the Mac Pro. Of course that begs the question too of, what exactly would be user upgrade-able inside of a machine that is basically 100% proprietary - and that is in fact a very interesting question - but in theory if they allowed for CPU/GPU upgrades (even if only directly through Apple themselves) that would/could increase the longevity of the machines and still perhaps keep standardized slots for PCI-E and RAM - which would lower the cost of those components not only just for the user base but also Apple themselves.

As another wrench, I think Apple also wants to be a premiere machine in the graphics space. It's little known, even to Apple fans, but Metal has API's already in it for both RT and VR. Apple launched both of those API's over 3 years ago. They were actually ahead of the curve at least in the sense of ensuring that those things could be coded for on macOS. However they've been lacking the hardware support. Obviously AMD only recently has cards that can support RTX features and there isn't really any game development or VR happening on macOS - but the intention again is to capture that space. Apple is going to have to do a heck ton to catch up there too if they want to attract devs and consumers. Not having RT is going to matter more and more and it's not something Apple can ignore for long. Not even in their first gen product. In 2 years when the Mac Pro launches if it doesn't have RT that 's going to be woefully behind. That will likely be 3 RT generations for nVidia and at least 2 for AMD. To not have it on a professional level machine will just be a non-starter for basically all devs and it won't make sense at all versus just buying and building a WinTel machine.

That's why I think at least short term they're going to have to have an AMD discrete GPU. I'm not sure there is any logical way around it.
LOL. Agreed. Yes, hopefully I made that clear enough in the post. Linear scaling estimates are sort of like sticking your thumb in the air to get a weather report. In this case, it's more about scoping the problem. Scaling from the M1 8c GPU to the rumored 16-24c likely is fairly close to the linear estimate, but that's only going to work for the laptops and other relatively low power devices. Scaling it 16x to 128c is not at all the same.

If a linearly scaled 128c GPU (cores only) needs about 400mm^2, without all the memory controllers, interconnects, cache (?), advanced features a workstation requires, etc. then I don't see how they can make a SoC with that. The rumored 32c CPU plus that is way over the reticle limit of TSMC 5nm. Not to mention, scaling to 32c CPUs isn't exactly a cut and paste job either! They'll need to go with a discrete option for the really big stuff. Now, that could be some cool multi-chip module, a monolithic die just isn't in the cards.

I know Nvidia gets hammered about spending so much die space on RT cores, but I agree, it will be a killer feature when it finally is good enough and we have software to make use of it. And you have good points about where things will be whenever they get around to launching a large GPU. Two things. First, like you said, they have the Metal framework support, so it's likely they've been playing with something for a while. We just don't know what because nothing released has iTracing yet. Second, it's not RT, but Apple does have the Afterburner FPGA IP. Intel's been working similar programmable hardware angles for a while now and AMDs onboard as well. Will be interesting to see how that all plays out.

I bring it up Afterburner because it's a good example of what Apple can do when they throw a lot of resources at a problem they really care about, even if in the traditional MBA sense, it may not make sense* (I still don't know how they recovered the NRE for Afterburner, even at the prices Apple charges). Let's not kid ourselves about scale here. Apple has cash on hand that's greater than Nvidia's revenues for the last decade. Apple's gross annual profit is an order of magnitude larger than Nvidia's revenue. Let that sink in for a moment. Then consider that Apple's revenue is almost 4x that of Intel's! Now money isn't everything, but it does help if you decide you really want to solve a specific problem and can apply laser like focus. Perhaps you're right and the discrete systems will go through a whole generation, rumored to be approx. 2yr cycles(?), before Apple's ready to even try taking on AMD/Nvidia.

I personally think the writing is on the wall for anything mobile. The performance is there and the power savings is tremendous and the profit margin is awesome. I really can't see Apple using AMD in a 16" MBP. The iMac Pro and Mac Pro, that's a bit of different story. However, I think they may have been working on this for a lot longer than assumed. Should be fun to watch.


Edit: * I mean this in the short term ROI sense. I think it does make sense in as part of a longer term, strategic move.
 
Top